Healthcare Compliance

healthcare-compliance checks model answers for medical phrases. It blocks an answer that contains a phrase you list, such as "stop taking", and puts a disclaimer in front of an answer that looks like medical advice.

It does not check that clinical information is correct, and it does not make you compliant with a regulation. fda_class only selects a built-in disclaimer. It does not classify or approve your product. The control runs only with stream: false.

Configuration

Save this as policy-config.yaml and apply it to a test agent with the config-first workflow.

policy-config.yaml

pack:
  name: healthcare-compliance-example
  version: 1.0.0
  enabled: true
policies:
  chain:
    - healthcare-compliance
policy:
  healthcare-compliance:
    blocked_patterns:
      - diagnose you with
      - prescribe
      - stop taking
      - surgery is necessary
    required_disclaimers:
      - This is not medical advice.
      - Consult a licensed clinician for diagnosis or treatment decisions.
    fda_class: II

Verify the result

Use controlled provider responses, because model answers vary:

Provider response	Result
I cannot prescribe medication.	Blocked with healthcare.blocked. The phrase matches even in a refusal.
The treaty was signed.	Disclaimer added with healthcare.disclaimer_injected, because treat matches
Hello.	Returned unchanged with healthcare.clean

Compare the client response with the nested result in the runtime event, and tune the phrases with test data from your application before rollout.

Fields

Field	Type	Default	Description
blocked_patterns	list of strings	[]	Literal substrings that block the answer. Case does not matter. Not regular expressions.
required_disclaimers	list of strings	[]	Lines joined with line breaks and put before an answer that looks like medical advice. Empty uses the built-in disclaimer of fda_class.
fda_class	string	II	I, II, or III. Selects the built-in disclaimer and appears in the details. Adds no block rules.

fda_class	Built-in disclaimer
I	General health information only; not medical advice.
II	This is not medical advice. Consult a licensed clinician.
III	This is not medical advice and must not be used for diagnosis or treatment decisions. Consult a licensed clinician.

How it works

1. The gateway takes the text of the complete model response.
2. The control lowercases it and looks for each blocked_patterns entry. The first match blocks with healthcare.blocked and records it in matched_pattern.
3. Otherwise, the text counts as medical advice if it contains diagnose, diagnosis, treat, treatment, prescribe, prescription, dosage, dose, start taking, stop taking, contraindication, or side effects, or take with mg or ml, or a dose phrase such as take 500 mg or increase 200 mcg.
4. Advice that does not start with the disclaimer gets it, with two line breaks, and healthcare.disclaimer_injected.
5. Advice that already starts with it gets healthcare.disclaimer_present.
6. Anything else is returned with healthcare.clean.

The details contain fda_class, looks_like_medical_advice, and disclaimer_injected, or for a block, blocked: true, matched_pattern, and fda_class.

Phase and verdicts

Item	Value
Phase	output
Request families	Chat Completions, Completions, Responses, Messages
Verdicts	allow, block
Reason codes	healthcare.clean, healthcare.disclaimer_present, healthcare.disclaimer_injected, healthcare.blocked

Examples

Block prescribing language with your own disclaimer:

Inside policy-config.yaml

policy:
  healthcare-compliance:
    blocked_patterns:
      - prescribe
      - increase the dose
      - stop taking
    required_disclaimers:
      - This is not medical advice.

Use the class III disclaimer:

Inside policy-config.yaml

policy:
  healthcare-compliance:
    blocked_patterns:
      - diagnose you with
      - this is cancer
    fda_class: III

Limits and failure behavior

• A block returns HTTP 400 with error code content_policy_violation. The client gets no text.
• The advice check matches substrings: treat matches treaty and retreat, and dose matches endorse.
• Only blocked_patterns blocks. The advice check and fda_class only add disclaimers.
• blocked_patterns also matches in longer words and in quoted or negated text: prescribe matches I cannot prescribe medication.
• The control compares the start of the answer, after leading spaces, with your full disclaimer. An answer that starts with a different disclaimer still gets yours.
• Lint accepts only I, II, or III for fda_class. The runtime accepts any string and uses the class II disclaimer for an unknown value.
• With stream: true, the gateway buffers the response but does not run the control.
• A when condition does not limit this control. Use routes.

Next steps

• HIPAA PHI Detector: health information in requests.
• Financial Compliance: the same pattern for investment advice.
• Investigate a Blocked Request: trace a healthcare.blocked decision.